A time series problem in telecommunications: Physical Resource Block (PRB) forecasting

Time series forecasting holds a crucial role in any industrial or institutional context since the knowledge and the analysis of historical data allow organizations to make informed decisions and optimize processes. Accurate predictions are the foundation for valuable business strategy since they could result in additional revenues or cost savings. Therefore, the integration of Machine Learning modules into forecasting tasks and the continuous exploration and development of novel and domain-specific models represent areas of extensive and ongoing investigation. This work explores some of these tools as part of a project of energy saving realized by Spindox Spa and committed by an important telecommunication company. The provided dataset consists of different data flow information, including Physical Resource Block (PRB), hourly collected for about 40 days in a cellular network, and it is used to predict, for each of the cells, the PRB stream during the following 24 hours for each cell. We explore two distinct approaches and evaluate the quality of the results by forecasting error, measured with the Weighted Mean Absolute Percentage Error (WMAPE) metric. First, we design a gradient-boosting framework using LightGBM Regressor, a powerful machine learning model known for reaching remarkable performance in terms of accuracy, speed and efficiency. Through the selection of suitable regressors and an optimized search of the best configurations for the model, we achieve successful outcomes. The obtained WMAPE scores are pretty favorable, particularly given the complexity of the problem. The second approach involves Functional Data Analysis (FDA), a statistical technique that employs theoretical tools of functional analysis in order to understand complex temporal patterns in the data and enhance the precision of the forecasting process. In FDA, the data are functions observed, possibly with error, over the time domain. In our work, we consider day-functions with 24-hour domain, for each cell and each available day, and we build a forecasting model based on Functional Principal Component Analysis to predict the function relating to the next day. Considering the limitedness of current tools and resources, the obtained results should be satisfactory and a promising starting point for future study. Finally, we compare the two approaches and outline some ideas for further research.